Diagnosis 2020; 7(1): 3–9
Mini Review
Paul A. Bergl*, Thilan P. Wijesekera, Najlla Nassery and Karen S. Cosby
Controversies in diagnosis: contemporary debates
in the diagnostic safety literature
https://doi.org/10.1515/dx-2019-0016
Received March 1, 2019; accepted April 28, 2019; ­
previously
­published online May 27, 2019
Abstract: Since the 2015 publication of the National Acad-
emy of Medicine’s (NAM) Improving Diagnosis in Health
Care (Improving Diagnosis in Health Care. In: Balogh
EP, Miller BT, Ball JR, editors. Improving Diagnosis in
Health Care. Washington (DC): National Academies Press,
2015.), literature in diagnostic safety has grown rapidly.
This update was presented at the annual international
meeting of the Society to Improve Diagnosis in Medi-
cine (SIDM). We focused our literature search on articles
published between 2016 and 2018 using keywords in
Pubmed and the Agency for Healthcare Research and
Quality (AHRQ)’s Patient Safety Network’s running bib-
liography of diagnostic error literature (Diagnostic Errors
Patient Safety Network: Agency for Healthcare Research
and Quality; Available from: https://psnet.ahrq.gov/
search?topic=Diagnostic-Errors&f_topicIDs=407). Three
key topics emerged from our review of recent abstracts
in diagnostic safety. First, definitions of diagnostic error
and related concepts are evolving since the NAM’s report.
Second, medical educators are grappling with new
approaches to teaching clinical reasoning and diagno-
sis. Finally, the potential of artificial intelligence (AI) to
advance diagnostic excellence is coming to fruition. Here
we present contemporary debates around these three top-
ics in a pro/con format.
*Corresponding author: Paul A. Bergl, MD, Assistant Professor
of Medicine in the Division of Pulmonary, Critical Care, and Sleep
Medicine, Froedtert and the Medical College of Wisconsin, Hub for
Collaborative Medicine, 8th Floor, 8701 W. Watertown Plank Road,
Milwaukee, WI 53226, USA, Phone: +414-955-7047,
E-mail: pbergl@mcw.edu. https://orcid.org/0000-0002-9406-2792
Thilan P. Wijesekera: Section of General Internal Medicine,
Yale School of Medicine, New Haven, CT, USA,
E-mail: thilan.wijesekera@yale.edu
Najlla Nassery: Division of General Internal Medicine, Johns
Hopkins University School of Medicine, Baltimore, MD, USA,
E-mail: nnasser3@jhmi.edu
Karen S. Cosby: Department of Emergency Medicine, Rush Medical
College, Chicago, IL, USA, E-mail: kcosby40@gmail.com
Keywords: artificial intelligence; clinical reasoning;
­diagnostic error; machine learning; management reason-
ing; overdiagnosis; patient safety.
Position 1 – The 2015 National
Academy of Medicine definition
of diagnostic error [1] best serves
­clinicians and patients
PRO
In 2015, the National Academy of Medicine (NAM)
defined diagnostic error as a “failure to (a) establish an
accurate and timely explanation of the patient’s health
problem(s) or (b) communicate that explanation to the
patient” [1]. In the process of developing this definition,
the NAM substantially advanced our conception of “diag-
nostic errors” from the limited scope conveyed in To Err
is Human [2] and the clinician-oriented definitions popu-
larized in the early 21st century [3, 4]. While earlier defini-
tions situated diagnostic error within healthcare systems
[2, 4], the NAM advanced a definition that accounts for
how teams and their ecosystems influence the diagnos-
tic process (Figure 1). In addition, the NAM model seem-
ingly reconciled debates about which diagnostic errors
really matter [5] and whether processes of arriving at a
diagnosis – i.e. the “workup” – or disease labels – i.e.
what medical terms we use to codify the patient’s symp-
toms and signs – are of higher importance [6]. The NAM
definition also accounted for two increasingly appreci-
ated aspects of the diagnostic process: uncertainty and
patient-centeredness.
Moving away from narrow, black-and-white con-
cepts like “delayed diagnosis” and “misdiagnosis”,
the first clause of the NAM definition appreciates the
nuance and challenge of diagnostic uncertainty [7, 8].
Often, an unconfirmed hypothesis is a clinician’s most
accurate explanation for a patient’s symptoms. Under
the NAM approach, these uncertainties do not represent
4      Bergl et al.: Contemporary debates in diagnostic safety

Dat
ai
nte
on rp
ati
gr

re
tat
e
New health

int

ion
problem

ta
Da
Clinical Physical
history examination
Patient
engagement Communication
Treatment Outcomes
with healthcare of diagnosis
Referral and Diagnostic
system consultation testing

on
a ti
I nf

er
or

at

en
m

io g
n s
ga se
th e he
r in g Hy p ot

Response to treatment

Figure 1: The diagnostic process, adapted from the NAM model [1].

unsettling failures, but rather byproducts of the iterative
nature of the diagnostic process (Figure 1) [1]. Prioritiz-
ing explanations over diagnostic certainties might not
only improve diagnostic decision making but could also
reduce provider burnout and engender higher trust with
patients [7, 9, 10].
The NAM definition and model also reinforce the
importance of teamwork in the diagnostic process, an
imperative given the complexity of the healthcare system.
More importantly, the NAM definition includes the patient
on the diagnostic team [1]. Arguably, because of the NAM’s
model, diagnostic teamwork and patient-centeredness are
gaining more attention than ever before [11]. The NAM
definition calls us to move beyond the obvious benefits
of interdisciplinary diagnosis, such as improved com-
munication amongst clinicians [12, 13], and it aligns with
larger trends in healthcare to incorporate the patient into
­team-based care [14].
CON
The NAM’s definition falls short of the greater goal of
improving diagnosis in healthcare. Because the defini-
tion of diagnostic error is not risk-adjusted, all errors are
treated equally, whether or not they result in permanent
morbidity or mortality. In addition, the NAM definition
prioritizes eliminating underdiagnosis without recogniz-
ing the risks of overdiagnosis. Finally, the NAM’s approach
does not repudiate flawed diagnostic processes that even-
tually achieve accurate diagnostic labels despite inflicting
collateral damage, such as excessive testing or complica-
tions from inappropriate procedures.
The diagnostic error community has struggled with
how to handle diagnostic errors that do not harm patients
[15]. Misdiagnoses around benign, self-limited conditions,
such as confusing viral bronchitis with bacterial sinusi-
tis, do not substantially impact patient outcomes though
they represent diagnostic errors. While these errors may
expose other cognitive or procedural failures, they detract
attention from the misdiagnoses that are most amenable
to practical solutions and most injurious to patients, like
a missed epidural abscess or bacterial meningitis [16].
Using the framework of “undesirable diagnostic events”
[16], rather than the NAM definition, the diagnostic safety
community would focus on readily diagnosable condi-
tions that satisfy the following criteria: (1) The diagnosis
has objective, valid reference standards; (2) The diagnosis
is prone to error; (3) The diagnosis benefits from timely
recognition; and (4) There are measurable, preferred
pathways to achieve diagnosis [16]. Table 1 expands upon
two commonly missed diagnoses that align with this
approach.
Although mentioned in the NAM report, overdiag-
nosis is not incorporated into the definition of diagnos-
tic error [1]. In its strictest sense, overdiagnosis refers to
detection of a disease that would have never caused harm
in a patient’s lifetime; indolent cancers detected only
through screening tests represent a classic example [17].
Related problems include (1) overdetection of irrelevant
findings, such as thyroid nodules found on computed
tomography (CT) scans ordered for suspected pulmonary
 Bergl et al.: Contemporary debates in diagnostic safety      5

Table 1: Two examples of commonly missed diagnosis that may represent undesirable diagnostic events.

Presenting    Diagnosis   Objective   Benefit(s) of early recognition   Preferred pathways to diagnosis

complaint standard

Incidental    Lung   Biopsy   Lower stage at diagnosis (if cancer),   – Dedicated lung imaging (CT) for suspicious nodules
lung nodule cancer longer survival – Guidelines for evaluation based on nodule size
– Multiple approaches to obtain tissue (biopsy)
Nausea and    MI   High-sensitivity   Lower risk of death and complications   – 12-lead ECG
abdominal troponin assays, including enlarging territory of – Troponin
pain consensus infarcted myocardium, heart failure, – ECG and/or cardiac MRI
definitions of MI apical thrombus, etc. – Coronary angiography

CT, computed tomography; ECG, echocardiography; MI, myocardial infarction; MRI, magnetic resonance imaging.

embolism; (2) expanding disease definitions, such as Position 2 – Medical education in

decreasing thresholds of glycated hemoglobin to diagnose
diabetes mellitus; and (3) overmedicalization of poten-
tially normal variants, the so-called “disease mongering”
that gave rise to entities like testosterone deficiency and
restless legs syndrome [17]. Clinicians constantly face the
tenuous balance of promptly recognizing and treating
disease against the risks of harm from over-testing and
overtreatment. Common conditions like sepsis exemplify
the importance of maintaining over- and underdiagno-
sis in equilibrium [18]. When sepsis is under-recognized,
more patients succumb to organ dysfunction and death.
Yet if every hospitalized patient with hypotension were
evaluated and treated for suspected sepsis, resource uti-
lization would skyrocket. In addition, patients would
suffer tangible consequences like antibiotic-related side
effects and fluid overload. The NAM definition’s focus on
reducing underdiagnosis oversimplifies the goals of diag-
nostic safety.
Highlighting undesirable diagnostic events and
overdiagnosis align with larger goals of improving the
diagnostic process, which includes cognitive steps by
the clinician and clinical activities within the broader
system [19]. As both are amenable to continuous process
improvement and educational interventions [19], they
may be even more relevant than NAM-defined diagnos-
tic errors, even in the absence of overt misdiagnosis.
Consider the example of pulmonary embolism. While
routinely ordering CT pulmonary angiography for every
patient suspected of having a pulmonary embolism
would reduce rates of missed diagnosis, such an evalua-
tion carries risks of radiation exposure, false-positives,
incidental findings, and high cost. Improved diagnos-
tic processes, such as integrating pulmonary embolism
risk prediction calculators into decision support tools,
could result in more efficient evaluations and greater
cost-effectiveness, thereby promoting homeostasis
between over- and underdiagnosis [19].
clinical reasoning should address
diagnostic and management
reasoning as related albeit distinct
concepts
PRO
Though an undeniably important part of the diagnos-
tic process, clinical reasoning remains challenging to
define [19]. One recent suggestion has been to distinguish
between diagnostic reasoning and management reason-
ing in clinical reasoning pedagogy [20]. Management
reasoning has been defined as “the process of making
decisions about patient management, including choices
about treatment, follow-up visits, further testing, and
allocation of resources [20].” It differs from diagnostic
reasoning in that management reasoning can have multi-
ple correct answers, is influenced by clinician and patient
preferences, requires shared decision-making, is intrinsi-
cally dynamic, and is dependent on context. On the con-
trary, diagnostic reasoning largely involves classifying
clinical findings into useful labels, such as syndromes
and diseases.
While diagnostic reasoning and management reason-
ing often occur in parallel, prioritizing the latter confers
several advantages to improving trainees’ diagnostic
acumen [21]. First, whereas the cognitive processes of
arriving at a diagnosis are difficult to deconstruct and
measure, the analytical and decision-oriented nature of
management reasoning provides more discrete opportuni-
ties for education and assessment [20]. Second, manage-
ment reasoning can play a role in navigating diagnostic
uncertainty, a challenging topic to teach learners [7, 22].
When a diagnosis is unclear, educators can still assess
6      Bergl et al.: Contemporary debates in diagnostic safety

a learner’s ability to create a management plan, includ-
ing judicious ordering of tests, collaborating with other
health care providers, and writing accurate documenta-
tion [20]. Third, management reasoning entails questions
such as whether a diagnosis should even be pursued or
what changes in management will come from diagnostic
testing. Thus, it aligns with broader efforts to promote
high-value care in medical education [23, 24].
Testing and treatment thresholds represent an
underutilized approach to teaching management reason-
ing (Figure 2A) [25]. Though the concept was introduced
nearly 40  years ago, students still have only a “fair”
understanding of thresholds, likely due to under-
emphasis of this concept in medical training [26]. While
diagnostic probabilities are encapsulated in any thresh-
old-based management decisions, this approach to
management reasoning focuses on practical, situational
decision-making that is largely independent of the diag-
nostic process. Providing a clear structure for manage-
ment reasoning might include delineating the types of
management (e.g. medications, laboratory studies) and
parties involved (i.e. patient, clinician, and system). For

Figure 2: The test-treatment threshold approach to clinical decision-making (A) and factors and dimensions that modify thresholds (B) as
adapted from Pauker and Kassirer [25].
In Panel A, diagnoses considered unlikely fall below the testing threshold (Tt) and do not merit additional diagnostic evaluation. Diagnoses
considered likely enough that the benefits of treatment outweigh the risks of awaiting diagnostic testing are above the treatment threshold
(Ttx) and are empirically treated. As shown in Panel B, multiple factors and dimensions of a case may modify these thresholds.

example, deciding to prescribe antibiotics in a patient
with an upper respiratory infection might depend on the
acuity of illness, the patient preference to take antibiot-
ics, the provider comfort with prescribing antibiotics,
and the cost to the system from increasing antimicrobial
resistance (Figure 2B).
CON
Although separating these processes may have philosophi-
cal underpinnings, the diagnostic process and patient
management decisions are inextricably linked for prac-
ticing clinicians. Narrowly conceiving of diagnosis as a
“means to an end” for proper management [20] or an exer-
cise in disease labeling allows for this artificial separation.
This simplistic approach, however, belies the complexity
of the diagnosis process. Instead, clinicians often use the
diagnostic process as part of a larger, more dynamic mental
framework for making sense of clinical data [27]. Again,
take the evaluation of a hypotensive patient with suspected
sepsis as an example. On initial evaluation, the diagnostic
hypothesis of sepsis is rarely certain. As the patient’s illness
evolves, the probability of sepsis is refined through man-
agement, whether watchful waiting, evaluation for compet-
ing alternative diagnoses, or clinical responses to empiric
treatments such as fluid loading and antibiotics. As clini-
cians observe this evolution, they re-evaluate treatment
strategies and iteratively reprioritize diagnoses. Ergo, diag-
nosis and management are fluid, co-dependent processes;
diagnostic reasoning is management reasoning.
Creating a false dichotomy threatens advances in both
diagnostic excellence and management reasoning. Many
prioritized areas in diagnostic safety, such as overdiagnosis,
handling uncertainty, and diagnostic stewardship, reflect
the codependence of diagnosis and management. For
example, overdiagnosis occurs when a legitimate medical
condition that would have not caused symptoms or harm
is discovered through aggressive case finding [28]. As dis-
cussed, overdiagnosis takes many forms, but all examples
cited earlier, such as early cancer detection or shifting gly-
cated hemoglobin thresholds, reflect the tension of impact-
ing the disease course by diagnosing it earlier against the
harms of early detection – i.e. management reasoning. Sim-
ilarly, the decision to even label a newly discovered disease
inherently includes value judgments about potential man-
agement. For example, as genome and exome sequencing
become commonplace, we will continually face questions
about whether variants represent true disease and how to
act upon such results [29]. Finally, as diagnostic stewards,
clinicians can only justify testing when they understand
the therapeutic implications, and therapy is defensible
Bergl et al.: Contemporary debates in diagnostic safety      7
only when we understand how diagnostic testing informs
disease probability. Using the contemporary prototype
of diagnostic stewardship Clostridioides difficile (C. diff.)
testing, clinicians should not only consider whether a
patient’s clinical syndrome likely represents true colitis
before testing but also whether a positive test result would
prompt treatment [30].
In other words, clinicians constantly face test-treat-
ment thresholds. Accordingly, clinicians must weigh the
reliability of diagnostic tests, the risks of overtesting, and
implications of treatment [25]. The false divide between
diagnostic reasoning and management reasoning detracts
from the broader goals of improving critical thinking.
Promising pedagogical approaches, such as teaching
relational reasoning, can potentially unify these related
topics [31]. Relational reasoning prompts learners to seek
meaningful patterns in clinical data, a skillset suited for
both diagnosis and treatment. Failing to recognize and
teach the inter-related nature of diagnosis and treatment,
however, is a threat to both fields.
Position 3 – Artificial intelligence
is a distraction, not a panacea. The
future of diagnostic medicine rests
in human hands
PRO
In recent years, artificial intelligence (AI) has shown
promise in automating steps in the diagnostic process.
For example, using machine learning, algorithms can flag
potential cases of intracranial hemorrhage on CT [32] or
identify patients at risk of clinical heart failure based on
echocardiographic computations [33]. These advances
could expedite care and prioritize cases requiring expert
clinician review, thereby revolutionizing how we arrive at
diagnoses.
Before debating the utility of AI in diagnostic safety,
however, one must understand how the machines actually
learn. Generally, two approaches predominate: supervised
and unsupervised learning [34]. With supervised learning
of diagnoses, the computer is provided with a set of vari-
ables which include a clinician’s determination of the diag-
nosis. Using the example of intracranial hemorrhage, the
machine reviews a large number of CT images and links
imaging patterns to the diagnoses provided by radiologists.
In unsupervised learning, the computer infers groups of
similar patients simply by associations of patient variables.
8      Bergl et al.: Contemporary debates in diagnostic safety
Thus, to participate in the diagnostic process, the
computer requires human oversight and quantifiable,
discreet variables. Supervised learning excels with a
human-derived reference, i.e. the gold standard of an
expert clinician’s diagnosis. The output of unsupervised
learning requires humans to contextualize and interpret
associations. Either way, the entire premise of using AI to
advance diagnostic excellence hinges on the availability
of expert human judgment [34]. Variables used in current
machine learning algorithms are based on numerical data
or discreet text inputs found in electronic data sets. Thus,
algorithms can handle imaging studies, which can be
­distilled into pixels, voxels, and other discreet quanta, but
may not process nuances or irregularities.
Reliance on AI could have other undesirable effects in
diagnostic safety. Most saliently, progressive dependence
on diagnostic technology can perpetuate “de-skilling”, or
the attrition in human skills [35]. Given that AI requires
human inputs for optimization, this de-skilling presents
an enormous threat to diagnostic accuracy. Arguably, our
collective love affair with technology has already created
an environment in which we neglect the basic steps in
the diagnostic process, such as history-taking or physical
examination. These basic failures account for many diag-
nostic errors in contemporary medicine [36, 37]. Moreover,
AI is prone to bias; it lacks the insight to recognize anom-
alies or to grasp the importance of the non-randomness
of missing data [38]. The latter problem can magnify
health disparities and could leave out groups who tradi-
tionally have poor access to care. An abundance of data
from majority ethnic groups or high socioeconomic status
patients further risks biasing AI as the computer will learn
diagnostic patterns from these over-represented groups.
CON
Homo sapiens have progressed as a species by innovating
and adapting to change. We are fortunate to live in an era
of immense technological capacity; we have access to large
clinical data sets and can use advanced computational
methods, namely AI and machine learning, to optimize the
diagnostic process. Ignoring these tools in healthcare risks
stagnation and would be a disservice to our patients.
AI and human cognition have different strengths,
and few believe AI to be a panacea for diagnostic safety.
With proper oversight, however, AI can overcome the
limitations of human cognition that have become abun-
dantly clear in the information age. A single person or
group of persons cannot efficiently collate, internalize,
and analyze large – let alone massive – data sets. Using
incredible computing power, machine learning can iden-
tify patterns among these data, integrate seemingly dispa-
rate data points, and even make sense of healthcare data
in novel ways [39]. In other words, machine learning will
not only expedite clinicians’ abilities to recognize patterns
in data for individual patients, but it will likely uncover
and define entirely new diagnoses.
The computational power of AI will enhance diagnos-
tic safety in other novel ways. The connectivity and upda-
tability of computer networks give them another edge
on human decision-makers who have difficulty keeping
current and assimilating new knowledge [40]. As such,
AI can rapidly integrate new scientific discoveries in diag-
nostic algorithms. In addition, AI will advance our under-
standing of diagnostic errors [41]. Harnessing big data
will allow us to view interconnected data points, such as
patterns of patient visits or symptom-disease associations
[42], which may signal potential delays in diagnosis. In
this way, AI will provide another avenue for investigating
the magnitude and causes of such errors.
AI itself is a product of human knowledge, and
machine learning is a “natural extension [of] traditional
statistical approaches” [39]. Similarly, we must view
AI in our clinical work and research as expected evolu-
tions in science. As we humans continue to expand our
understanding of health and disease, we will work closely
alongside AI-guided computers. We will program them,
refining their algorithms, and accept or reject their sug-
gestions, giving them the necessary feedback to improve.
We must embrace this synergy.
Author contributions: All the authors have accepted
responsibility for the entire content of this submitted
manuscript and approved submission.
Research funding: None declared.
Employment or leadership: None declared.
Honorarium: None declared.
Competing interests: The funding organization(s) played
no role in the study design; in the collection, analysis, and
interpretation of data; in the writing of the report; or in the
decision to submit the report for publication.
Disclosures: The authors have no relevant conflicts to
disclose.
References
1. Improving Diagnosis in Health Care. In: Balogh EP, Miller BT, Ball
JR, editors. Improving Diagnosis in Health Care. Washington (DC):
National Academies Press, 2015.

2. Kohn LT, Corrigan JM, Donaldson MS, editors. To err is human:
building a safer health system. Washington (DC): National Acad-
emies Press, 2000.
3. Graber ML, Franklin N, Gordon R. Diagnostic error in internal
medicine. Arch Intern Med 2005;165:1493–9.
4. Schiff GD, Kim S, Abrams R, Cosby K, Lambert BL, Elstein AS,
et al. Diagnostic error in medicine: analysis of 583 physician-
reported errors. Arch Intern Med 2009;169:1881–7.
5. Singh H. Editorial: helping health care organizations to define
diagnostic errors as missed opportunities in diagnosis. Jt Comm
J Qual Patient Saf 2014;40:99–101.
6. Newman-Toker DE. A unified conceptual model for diagnostic
errors: underdiagnosis, overdiagnosis, and misdiagnosis. Diag-
nosis (Berl) 2014;1:43–8.
7. Simpkin AL, Schwartzstein RM. Tolerating uncertainty – The next
medical revolution? N Engl J Med 2016;375:1713–5.
8. Bhise V, Rajan SS, Sittig DF, Morgan RO, Chaudhary P, Singh H.
Defining and measuring diagnostic uncertainty in medicine: a
systematic review. J Gen Intern Med 2018;33:103–15.
9. Armstrong K. If you can’t beat it, join it: uncertainty and trust in
medicine. Ann Intern Med 2018;168:818–9.
10. Bhise V, Meyer AN, Menon S, Singhal G, Street RL, Giardina TD,
et al. Patient perspectives on how physicians communicate
diagnostic uncertainty: an experimental vignette study. Int J
Qual Health Care 2018;30:2–8.
11. Graber ML, Rusz D, Jones ML, Farm-Franks D, Jones B, Cyr Gluck J,
et al. The new diagnostic team. Diagnosis (Berl) 2017;4:225–38.
12. Thomas DB, Newman-Toker DE. Diagnosis is a team sport –
partnering with allied health professionals to reduce diagnostic
errors. Diagnosis (Berl) 2016;3:49–59.
13. Dickerson EC, Alam HB, Brown RK, Stojanovska J, Davenport MS.
In-person communication between radiologists and acute care
surgeons leads to significant alterations in surgical decision
making. J Am Coll Radiol 2016;13:943–9.
14. Li J, Talari P, Kelly A, Latham B, Dotson S, Manning K, et al.
Interprofessional Teamwork Innovation Model (ITIM) to promote
communication and patient-centred, coordinated care. Br Med J
Qual Saf 2018;27:700–9.
15. Zwaan L, Singh H. The challenges in defining and measuring
diagnostic error. Diagnosis (Berl) 2015;2:97–103.
16. Olson AP, Graber ML, Singh H. Tracking progress in improving
diagnosis: a framework for defining undesirable diagnostic
events. J Gen Intern Med 2018;33:1187–91.
17. Carter SM, Rogers W, Heath I, Degeling C, Doust J, Barratt A. The
challenge of overdiagnosis begins with its definition. Br Med J
2015;350:h869.
18. Klompas M, Calandra T, Singer M. Antibiotics for sepsis-finding
the equilibrium. J Am Med Assoc 2018;320:1433–4.
19. Hautz WE. When I say… diagnostic error. Med Educ 2018 Jun 4.
doi: 10.1111/medu.13602 [Epub ahead of print].
20. Cook DA, Sherbino J, Durning SJ. Management reasoning:
beyond the diagnosis. J Am Med Assoc 2018;319:2267–8.
21. Cook DA, Sherbino J, Durning SJ. Diagnostic vs management
reasoning – reply. J Am Med Assoc 2018;320:1818–9.
22. Scott I, Gingerich A, Eva KW. Twelve tips for clinicians dealing
with uncertainty when assessing learners. Med Teach. 2018;1–7.
doi: 10.1080/0142159X.2018.1494381 [Epub ahead of print].
23. Cayea D, Tartaglia K, Pahwa A, Harrell H, Shaheen A, Lang VJ.
Current and optimal training in high-value care in the internal
Bergl et al.: Contemporary debates in diagnostic safety      9
medicine clerkship: a national curricular needs assessment.
Acad Med 2018;93:1511–6.
24. Levinson W, Kallewaard M, Bhatia RS, Wolfson D, Shortt S,
Kerr EA. ‘Choosing Wisely’: a growing international campaign.
Br Med J Qual Saf 2015;24:167–74.
25. Pauker SG, Kassirer JP. The threshold approach to clinical deci-
sion making. N Engl J Med 1980;302:1109–17.
26. Rencic J, Trowbridge Jr RL, Fagan M, Szauter K, Durning S.
Clinical reasoning education at US medical schools: results from
a National Survey of Internal Medicine Clerkship Directors. J Gen
Intern Med 2017;32:1242–6.
27. Ilgen JS, Eva KW, Regehr G. What’s in a label? Is diagnosis
the start or the end of clinical reasoning? J Gen Intern Med
2016;31:435–7.
28. Kale MS, Korenstein D. Overdiagnosis in primary care: framing
the problem and finding solutions. Br Med J 2018;362:k2820.
29. Splinter K, Adams DR, Bacino CA, Bellen HJ, Bernstein JA,
Cheatle-Jarvela AM, et al. Effect of genetic diagnosis on
patients with previously undiagnosed disease. N Engl J Med
2018;379:2131–9.
30. Morgan DJ, Malani P, Diekema DJ. Diagnostic stewardship-
leveraging the laboratory to improve antimicrobial use. J Am
Med Assoc 2017;318:607–8.
31. Dumas D, Torre DM, Durning SJ. Using relational reasoning
strategies to help improve clinical reasoning practice. Acad Med
2018;93:709–14.
32. Titano JJ, Badgeley M, Schefflein J, Pain M, Su A, Cai M, et al.
Automated deep-neural-network surveillance of cranial images
for acute neurologic events. Nat Med 2018;24:1337–41.
33. Tabassian M, Sunderji I, Erdei T, Sanchez-Martinez S, Degio-
vanni A, Marino P, et al. Diagnosis of heart failure with preserved
ejection fraction: machine learning of spatiotemporal varia-
tions in left ventricular deformation. J Am Soc Echocardiograp
2018;31:1272–84.e9.
34. Syeda-Mahmood T. Role of big data and machine learning
in diagnostic decision support in radiology. J Am Coll Radiol
2018;15(3 Pt B):569–76.
35. Cabitza F, Rasoini R, Gensini GF. Unintended consequences
of machine learning in medicine. J Am Med Assoc 2017;318:517–8.
36. Verghese A, Charlton B, Kassirer JP, Ramsey M, Ioannidis JP.
Inadequacies of physical examination as a cause of medical
errors and adverse events: a collection of vignettes. Am J Med
2015;128:1322–4.e3.
37. Clark BW, Derakhshan A, Desai SV. Diagnostic errors and
the bedside clinical examination. Med Clin North Am
2018;102:453–64.
38. Gianfrancesco MA, Tamang S, Yazdany J, Schmajuk G. Potential
biases in machine learning algorithms using electronic health
record data. J Am Med Assoc Intern Med 2018;178:1544–7.
39. Beam AL, Kohane IS. Big data and machine learning in health
care. J Am Med Assoc 2018;319:1317–8.
40. Harari YN. 21 Lessons for the 21st Century, 1st ed. New York City,
NY: Random House Publishing Group, 2018.
41. Dhaliwal G, Shojania KG. The data of diagnostic error: big, large
and small. Br Med J Qual Saf 2018;27:499–501.
42. Liberman AL, Newman-Toker DE. Symptom-Disease Pair Analysis
of Diagnostic Error (SPADE): a conceptual framework and meth-
odological approach for unearthing misdiagnosis-related harms
using big data. Br Med J Qual Saf 2018;27:557–66.